The CSIP Dataset
The Cross Sensor Iris and Periocular Dataset.
50 Participants
10 Mobile Setups
2000+ Images
+ Ground-Truth
When attempting to perform iris or periocular biometrics on mobile environments, several problems arise: 1) the wide variety of camera sensors and lenses mobile phones and tablets come equipped with produce discrepancies in working images, as they are acquired with color distortions, at multiple resolutions, etc.; 2) on-the-go acquisition by potentially untrained subjects will result in demanding Pose, Illumination and Expression (PIE) changes, as not all users hold their mobile devices at the same position, resulting in varying acquisition angles and scales, or rotated images; the acquisition environment can have poor or insufficient lighting, and uncontrolled outdoor daylight will most likely produce spectacle reflections over the iris region; 3) etc.
The main objective of the CSIP database was to gather images from a representative group of participants, acquired over cross-sensor setups and varying acquisition scenarios, thus mimicking the conditions faced on mobile application scenarios. Along with the data acquired with different mobile devices, an iris segmentation mask is also provided, allowing assessing the performance of both iris and periocular segmentation and recognition algorithms on mobile environments.
Mobile devices used for acquisition
Xperia Arc S
Sony Ericsson Xperia Arc S running Android 2.3.4. Images acquired using the standard camera application, with resolution 3264x2448, with and without flash.
Apple iPhone 4
Apple iPhone 4 running iOS 7.1. Images acquired using the standard camera application, resolutions 2592x1936 (rear) and 640x480 (frontal), with and without flash.
THL W200
THL W200 running Android 4.2.1. Images acquired using the standard camera application, resolutions 3264x2448 (rear) and 2593x1920 (frontal), with and without flash.
Huawei U8510
Huawei Ideos X3 (U8510) running Android 2.3.3. Images acquired using the standard camera application, resolutions 2048x1536 (rear) and 640x480 (frontal), without flash.
The Imaging Setup
Considering the heterogeneity of camera sensor/lens setups consumer mobile devices can deliver, a total of 10 different setups were used during the dataset acquisition stage: four different devices, some of them with frontal and rear cameras and LED flash. Aiming at mimicking the variability of noise factors associated with on-the-go recognition, participants were not imaged at a single particular location, but on multiple sites, as they were, with artificial, natural and mixed illumination conditions. As so, there is a substantial difference between each acquisition setup and surrounding conditions, even when the same setup was used to capture images from different subjects. From visual inspection, eight different noise factors are distinguishable, and can affect the biometric recognition process: multiple scales; chromatic distortions; image rotation; poor lighting; off-angle acquisition; out-of-focus images; deviated gaze; and iris obstructions (including reflexions).
Ground-Truth Data
For each periocular image acquired by the mobile devices, a binary iris segmentation mask is provided with the CSIP dataset. Those masks were automatically obtained using a state-of-the-art iris segmentation approach particularly suitable for uncontrolled acquisition conditions, which has been corroborated by the first place achieved at the Noisy Iris Challenge Evaluation - Part 1 (NICE.I)1.